Recursive Binding on a Budget: Subspace Carving in Order-p Tensor Memories
Researchers have introduced Orthogonal Subspace Carving (OSC), a novel memory architecture designed to enhance symbolic reasoning in AI models. OSC addresses the dimensionality issues of Tensor Product Representations by using projections to maintain a constant memory footprint, even for deep recursive structures. This approach allows for efficient binding of information within a fixed-size tensor, enabling component vectors to be significantly smaller than the memory tensor itself. AI
IMPACT Introduces a new memory architecture that could enable more efficient symbolic reasoning in AI models.